JPH0528323A - Character recognition device - Google Patents

Abstract

PURPOSE:To conduct an effective post-processing in accordance with the meaning of a document by selecting a resulted candidate character train (s) from one or more candidate word character train (s) based on the correlation with the word character trains. CONSTITUTION:A candidate word preparation part 130 combines the candidate character trains of respective word areas inputted through a word area cut-out part 120 and prepaes the word candidate character train (s). An appropriate candidate word is selected from these word candidate trains and outputted to a correlation detection part 142 in a result word selection part 140 as the candidate word. Discrimination of whether or not the word is appropriate as such word is based on whether or not the word candidate character train is present in a word dictionary 141 within the result word selection part 140 as an item. When no appropriate word can be obtained, the candidate word preparation part 130 outputs the word candidate character train prepared by combining the first candidate characters to the correlation detection part 142. When one word was outputted as the candidate word, a result word discrimination part 142 selects the candidate word as a resulted word.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device capable of recognizing characters with high accuracy.

[0002]

2. Description of the Related Art In recent years, a character recognition device capable of inputting a general document with high precision has been increasingly desired. This is for the purpose of advanced use of information and labor saving of information input. But,
When recognizing a general document written in Japanese, it is difficult to perform high-accuracy recognition only by character recognition, because homomorphic characters and similar-shaped characters exist in the document. Therefore, there has been proposed a character recognition device configured to perform post-processing using linguistic knowledge after recognizing each character.

As a character recognition device configured to perform such post-processing, for example, a document relating to the applicant of the present application (1988 Proceedings of the Spring National Convention of the Institute of Electronics, Information and Communication Engineers D-44) is proposed.
There is a character recognition device disclosed in 7). In this character recognition device, characters are recognized by the following processing.

First, a recognition process is performed on a recognition target character to obtain one or more candidate character strings. Next, the word area is detected by checking the connection of the same character type in this candidate character string. Next, the candidate character names in this word area are combined to create a word candidate character string. Next, it is determined whether or not these word candidate character strings are matched with the word dictionary to be a word. The word candidate character string determined to be established as a word is set as a candidate word. Then, when there is only one candidate word, the candidate word is set as a result word (meaning of the recognized word). When there are two or more candidate words, the average value of the similarity ranks of the candidate character names constituting each candidate word (hereinafter, referred to as “average candidate rank”) is compared, and the candidate word is the most. The candidate word having the highest average candidate rank is set as the result word.

[0005]

However, in the above-described conventional character recognition device, the evaluation value for selecting an appropriate word as a result word from a plurality of candidate words is determined by each of the candidate words. It was calculated based on the similarity rank of the candidate character names (for details, see column (D) of Table 1 below). Therefore, the character-by-character recognition result greatly affects the post-processing. Therefore, there is a problem that the post-processing does not function sufficiently as the post-processing. In particular, it is an input character string consisting of relatively simple characters, and there are some candidate characters for each character that make up this character string. However, combining these candidate characters yields many idioms. In the case of an input character string in which is extracted, there is a high possibility that the result word will be output according to the recognition result of each character. This will be described with a specific example.

For example, when the input character string is the "character" listed in the column (A) of Table 1 below, the "large letters" and " Three candidate character strings of "literature" and "jou" were obtained, and the candidate character names of these candidate character strings were combined to create the nine word candidate character strings listed in column (C) of Table 1 below. , The result of matching these word candidate character strings with the word dictionary Table 1 below
When the four candidate words “large letter”, “university”, “letter” and “literature” listed in column (D) are listed, each of the four candidate words is listed in column (D). The average candidate rank of is calculated. Then, "Large letters" with the smallest value was output as a result word, and this was not selected even though the correct word "letter" was listed as a candidate word, and eventually it was affected by the recognition result in character units. Candidate words are output.

As a result, the recognition accuracy is lowered, and a complicated operation such as a correction work is required to make up for it, which makes it impossible to efficiently input a Japanese document. It was

In order to avoid this, the post-processing unit is configured to enable higher-level language processing, for example, the mutual connection of words is tested by a syntactic analysis or semantic analysis method, and one of the candidate words is selected based on this result. A method of selecting a result word is also conceivable. However, in this case, since the language processing itself becomes complicated, a new problem arises such that the processing speed becomes slow and the device becomes large.

The present invention has been made in view of the above points, and therefore an object of the present invention is to provide a character recognition apparatus capable of easily performing effective post-processing according to the meaning of a document. It is in.

[0010]

To achieve this object, according to the character recognition device of the present invention, document image data is obtained from a document on a medium, and the character string in the document is obtained from the document image data. Character recognition unit that obtains a candidate character string in which each character is composed of one or more candidate characters, a unit character string cutout unit that cuts out an area corresponding to a unit character string from the candidate character string, and the above-mentioned cutout unit. Candidate unit character string creation unit that extracts one or more candidate unit character strings from the candidate character strings in each of the selected regions, and a result unit character that selects the resulting unit character string from the extracted candidate unit character strings In the character recognition device including the column selection unit, the result unit character string selection unit is a language dictionary in which predetermined unit character strings are associated with each other and registered, and the result unit character string selection unit was previously selected by the result unit character string selection unit. The simple result The unit character string storage unit for storing a character string, created by the candidate unit character string creation unit described above 1
A degree-of-association detection unit that detects the degree of association between the candidate unit character string and the resulting unit character string stored in the unit character string storage unit using the language dictionary described above; It is characterized by comprising a result unit character string selection unit that selects a result unit character string from the candidate unit character strings based on the degree of association detected by the above-described association degree detection unit.

[0011]

According to the structure of the present invention, unit character strings can be registered in the language dictionary in association with each other. Also,
The unit character string that is the resulting character string in the recognition target document is stored in the unit character string storage unit.

Then, the degree of association between one or more candidate unit character strings and the resulting unit character string stored in the unit character string storage unit is detected using the aforementioned language dictionary, and 1 is determined based on this degree of association. The resulting unit character string is selected from the above candidate unit character strings.

Here, the unit character string registered in the unit character string storage unit is included in the recognition target document which is the basis for extracting one or more candidate word character strings to be selected. It is a character string. Therefore, the character recognition device of the present invention, which selects a result candidate character string from one or more candidate word character strings based on the degree of association with such a word character string, is effective in conforming to the meaning of the document. It can be processed.

[0014]

Embodiments of the character recognition apparatus of the present invention will be described below. In the following examples, the unit character string will be described as an example of a word (including a character string in which the same character type is connected). Therefore, the unit character string cutout unit, the candidate unit character string creation unit according to the present invention, The result unit character string selection unit and the unit character string storage unit are referred to as a word area cutout unit, a candidate word creation unit, a result word selection unit and a word storage unit, respectively, and will be described. The language dictionary is a word dictionary in which predetermined words are registered in association with each other.

FIG. 1 is a block diagram schematically showing the configuration of a character recognition device 100 of the embodiment.

The character recognition device 100 of this embodiment includes a character recognition unit 110, a word region cutout unit 120 as a unit character string cutout unit, a candidate word creation unit 130 as a candidate unit character string creation unit, and a result character string selection unit. As a result, the word selection unit 140 and the output terminal 150 are provided.

In the case of this embodiment, the character recognition unit 110 of the character recognition apparatus 100 includes a photoelectric conversion unit 111, a line cutout unit 112, and a character cutout unit 1 as shown in FIG.
13, a sub-pattern extraction unit 114, a feature extraction unit 115, and a matching unit 116.

Further, the result word selection unit 140 of the character recognition device 100, in the case of this embodiment, as shown in FIG.
Word dictionary unit 141, relevance detection unit 142, word storage unit 1
43 and the result word determination unit 144.

Next, the details of each of the above-mentioned components will be described together with the description of the operation of the character recognition device 100. In order to facilitate this description, an example of a document image to be recognized will be described in the following description. FIG. 3 is a diagram showing a form used as a recognition target document in this embodiment. In FIG. 3, reference numeral 200 is a form, and 201 is a character entry frame. Further, Table 2 described later is a table for explaining how the input character string in the recognition target document shown in FIG. 3 is processed by the series of processes described below.

The optical signal S (see FIG. 1) from the form 200 as shown in FIG. 3 in which characters, figures, symbols and the like (hereinafter referred to as “characters”) are written is photoelectrically converted in the character recognition unit 110. It is input to the section 111.

The photoelectric conversion unit 111 photoelectrically converts the optical signal S, and quantizes it into an electric signal (for example, document image data) which is quantized into a binary black and white in which the character line portion is represented by black pixels and the background portion is represented by white pixels. Corresponding)) and outputs this electric signal to the line segmenting unit 112. The photoelectric conversion unit 111 is
For example, it can be configured by a conventionally known image sensor or the like.

The line segmenting section 112 is a photoelectric conversion section 111.
The input document image data is scanned with the character line direction as the main scanning direction and the character string direction as the sub-scanning direction, and a distribution of black pixels is created from this document image data. further,
In this black pixel distribution, the black pixels are "0" to "1".
From the position changing as described above to the position immediately before changing from “1” or more to “0” is cut out as character line image data, and the cut out character line image data is output to the character cutout unit 113.

The character cutout unit 113 is a line cutout unit 1.
The line image data input from 12 is scanned with the character line direction as the main scanning direction and the character string direction as the sub-scanning direction, and a distribution of black pixels is created from this line image data. further,
In this black pixel distribution, the black pixels are "0" to "1".
From the position changing as described above to the position immediately before changing from “1” or more to “0” is cut out as one-character pattern data, and this is output to the sub-pattern extraction unit 114.

In the case of this embodiment, the sub-pattern extraction unit 114 scans the character pattern input from the character cut-out unit 113 in a plurality of directions, and sets a predetermined value h (h in this embodiment) on each scanning line. = 5) A black pixel row in which black pixels are continuous is detected. Further, by extracting the black pixel row as a black pixel component of the sub-pattern, the sub-pattern for each scanning direction is extracted from the character pattern. Then, these sub-patterns are sequentially output to the feature extraction unit 115.

In the present embodiment, the plurality of scanning directions in the sub-pattern extracting section 114 are the directions (hereinafter, "X direction") perpendicular to the character line direction (hereinafter, "X direction"; see FIG. 3). Vertical direction), a direction parallel to the X direction (hereinafter referred to as "horizontal direction"), a direction 45 〓 counterclockwise from the X axis (hereinafter referred to as "left diagonal direction"), and from the X axis. The four directions are the 45 ° clockwise direction (hereinafter referred to as the “right diagonal direction”). Therefore, in the case of this embodiment, the sub-pattern extracting unit 114 extracts four sub-patterns of a vertical sub-pattern, a horizontal sub-pattern, a right diagonal sub-pattern and a left diagonal sub-pattern. Character pattern data is scanned with the main scanning direction as the vertical direction, and continuous black pixels (hereinafter referred to as “black runs”) are detected on this scanning line, and black runs having a continuous black pixel number L of L ≧ h are extracted. By doing so, the vertical sub-pattern can be extracted. Each of the horizontal, leftward, and rightward sub-patterns can be extracted in the same manner as the vertical sub-pattern extraction except that the main scanning direction is different.

In the case of this embodiment, the feature extraction unit 115
On each sub-pattern in each direction input from the sub-pattern extracting unit 114, a rectangular area corresponding to the character circumscribing frame of the character pattern is set, and this rectangular area is N × M (N and M are arbitrary suitable natural numbers). .) Small areas. Further, the length of the character line portion of each sub-pattern included in each small area is extracted as a feature amount for each small area of each sub-pattern. Further, these feature quantities are normalized by the size of the character circumscribing frame. Then, a feature matrix composed of the normalized feature quantities f _i is created for each sub-pattern in each direction, and these are output to the matching unit 116.

In the feature extraction unit 115 of this embodiment, the above-mentioned division numbers N and M are set to N = M = 8, and the above-mentioned normalization of the feature amount by the size of the character circumscribing frame is (dX + dY). /
The value shall be 2. However, dX is the horizontal length of the character circumscribing frame, and dY is the vertical length of the character circumscribing frame. Further, the feature quantity f _i is the number i (i = 1, 2, ..., N) from 1 to N × M in each of the N × M small areas.
XM) is a feature amount of the small area of number i when each small area is represented by being sequentially attached. The feature quantity f _i becomes the element value of the feature matrix.

The collation unit 116 collates the feature matrix F input from the feature extraction unit 115 with the feature matrix G of a standard pattern (not shown) prepared in advance, and the following (1)
The similarity R represented by the formula is obtained. Further, the character names of the dictionary matrix whose similarity R is equal to or greater than a predetermined value P are set as candidate character names, and the ranking is performed in order of the highest similarity, such as first candidate character and second candidate character. The one or more candidate characters are output to the word area cutout unit 120 as recognition result candidate characters.

[Equation 1]

[0029]

However, in the equation (1), g _i represents an element of the dictionary matrix.

The word area cutout unit 120 sequentially stores the candidate characters output from the collation unit 116 of the character recognition unit 110 in a storage unit (not shown) to create a candidate character string. In the case of the recognition symmetric document shown in FIG. 3, the word cutout unit 120 is based on the recognition result candidate characters extracted by the character recognition unit 110 for the input character string (refer to column (A) of Table 2 below) therein. Creates, for example, three types of candidate character strings shown in column (B) of Table 2 below.

Next, the word area cutout unit 120
A region corresponding to a word (this region is referred to as a "word region") is sequentially detected from the candidate character string created here (see column (B) of Table 2), and the candidate character string portion of the region is detected. Are sequentially output to the candidate word creation unit 130.

In the case of this embodiment, the word region is detected based on the character type of the first candidate character in the word candidate character string. That is, when detecting a word region, one word region is formed up to the point where a character of the same character type as the first candidate character of the first character of the detection target region becomes a first candidate character continuously. cut. For example, in the case of the table 2 (B) column, since the first character "pa" is katakana, the first word area is up to "pattern" in which the character type of katakana is the first candidate character. In such a word area cutout method, a hiragana area such as a case where particles are consecutive is detected as one word area, and thus a word area different from the original concept of a word may be cut out. Then, the region thus cut out may not be the target of post-processing performed later. However, in general, most of the words such as ordinary words or technical terms in a specific field are composed of a series of Chinese characters, a series of Katakana, a series of English characters, etc., so the word region used in this embodiment is used. The cutting method does not impair the effects of the invention.

From the candidate character strings listed in the column (B) of Table 2, the word region cutout unit 120 of this embodiment cuts out the word regions as shown in the column (C) of Table 2 and selects them. The words are sequentially output to the word creating unit 130.

The candidate word creating section 130 creates a word candidate character string by combining the candidate character strings of the respective word areas input from the word area cutting section 120, and further, a word that is more appropriate as a word than these word candidate character strings. Is selected, and the selected word candidate character string is output as a candidate word to the degree-of-association detecting unit 142 in the result word selecting unit 140.

In this embodiment, it is determined whether the word candidate character string is valid as a word in the word dictionary 141 (detailed description will be given later) of the word candidate character string in the result word selection unit 140. Depending on whether or not it exists as an item in. If no valid word is obtained, the candidate word creation unit 130 outputs a word candidate character string created by combining the first-ranked candidate characters to the degree-of-association detection unit 142. And

Considering an example in which the input character string is a "character" that is a part of the recognition target document shown in FIG. 3 (see the column (A) of Table 3 below), the word area cutout unit 12 is used.
0, the candidate word creation unit 130 creates candidate words as follows.

First, in the case of this embodiment, the word area cutout unit 120 extracts "large" from the input character string "character",
Candidate character strings “sentence”, “length” and “letter”, “gaku”, “U” are created (see column (B) of Table 3 below). Then, the candidate word creating unit 130 combines the candidate characters “large”, “sentence”, “length”, “letter”, “gaku” and “U” in these candidate character strings, and (C) in Table 3
Nine word candidate character strings shown in the column are created, and it is further determined whether or not these word candidate character strings exist as items in the word dictionary unit 141, and if they exist, the word candidate character strings are selected. Let it be a candidate word. (C) in Table 3
From the nine word candidate character strings in the column, four candidate words “large character”, “university”, “letter” and “literature” are selected as shown in column (D) of Table 3. These candidate words are output to the degree-of-association detection unit 142 in the result word selection unit 140.

In this embodiment, the word dictionary unit 141 has a structure that serves as both a word dictionary used in the later-described relevance test and a word dictionary used for selecting the above-mentioned candidate words. In order to facilitate the association test, the word dictionary 141 is a thesaurus-structured word dictionary in which each item is systematically classified and associated with each other. 4 and 5 are diagrams for explaining the word dictionary unit 141 of the embodiment. In particular, FIG. 4 shows the word dictionary unit 1 of the embodiment.
It is the figure which showed the structure of 41 notionally. In FIG.
The arrow indicates that the word at the start and the word at the end are related to each other. Further, in FIG. 4, the equal sign is 2
Indicates that two words are synonyms. Further, FIG. 5 shows a description example of a dictionary for "recognition" and "characters" in the document to be recognized. In FIG. 5, BT is a superordinate word of the word (a word such as “recognition” or “character”).
Also, NT indicates a lower word.

In order to facilitate the search, the word dictionary unit 141 is more preferably a dictionary unit having a map that allows each item to be searched by, for example, the first character and the number of characters. In addition, in this embodiment, the word dictionary unit 141 has a configuration that can be used for both the later-described relevance detection and the above-described word selection for determining whether or not the word is appropriate, but the present invention is not limited to this. However, the dictionary units may be separately prepared for the relevance detection and the word selection. In that case, the dictionary for testing whether or not the word candidate character string is valid as a word can have a simple structure in which only items can be searched. Further, the word dictionary unit may of course be provided in the candidate word creating unit 130.

The degree-of-association detecting unit 142 of the result word selecting unit 140 operates as follows according to the number of candidate words input from the candidate word creating unit 130 and the number of words stored in the word storage unit 143 described later. To do. The word storage unit 143 stores the word finally selected as the result word by the result word determination unit 144, and is easily realized by various storage media such as a memory, a flexible disk, a hard disk, and the like. it can.

First, when the number of candidate words input from the candidate word extraction unit is one, the candidate word is output to the result word determination unit 144 as it is.

On the other hand, when the number of candidate words is two or more and no word is stored in the word storage unit 143, the same relation is applied to all the candidate words input from the candidate word creation unit 130. A degree (for example, “1”) is added and output to the result word determination unit 144.

If there are two or more candidate words,
If any word is already stored in the word storage unit 143, each candidate word is compared with each of the words already selected as the result words stored in the word storage unit 143, and the degree of association is detected. For each candidate word, the average value of the degree of association with all of the stored words is calculated, and each candidate word and the calculated average degree of association are output to the result word determination unit 144.

The degree of association between the two words is represented by the distance between the two items, that is, the minimum number of branches connecting the two items in the word dictionary 141 having the thesaurus structure. The degree of association is stronger as the distance is smaller (the number of branches is smaller). For example, in FIG. 4, since the item “character” and the item “pattern” are directly connected by one branch (arrow), the degree of association is represented as “1”, and the item “university” and the item "Cognition" means "university"-"art"-"literature"-"book"-
Since “character”-“recognition” has been passed through five arrows, the degree of association is represented as 5.

For example, in the case of recognizing the input character string "character" illustrated in Table 3, as shown in the column of Table 3 (D), "large character", "university", "character" and "literature" are displayed. When two candidate words are extracted and when two words “pattern” and “recognition” are recognized as result words in the word storage unit 143 as in the column (D) of Table 2, The result of detecting the degree of association between each candidate word and each result word using the word dictionary as shown in FIG. 4 is as shown in column (E) of Table 3.

The method of detecting the degree of association, that is, the method of detecting the distance between items can be easily obtained by a conventional method using a dictionary having a description format as shown in FIG.

The result word determination unit 144 is the relevance detection unit 1.
Based on the candidate words input from 42 and the degree of relevance added to these candidate words, the candidate word most likely as the result word is selected as the final recognition result word from the candidate words and output to the output terminal 130. At the same time, whether or not the word is already stored in the word storage unit is tested, and if the word is not stored yet, the word is newly stored in the word storage unit 143.
Here, in this embodiment, the most probable candidate word as a result word is selected as follows.

First, when the candidate word output from the degree-of-association detecting unit 142 is 1, the result word determining unit 144 selects this candidate word as the result word.

On the other hand, when the number of candidate words input from the degree-of-association detecting unit 142 is two or more, the result word determining unit 144 determines that the candidate word having the highest degree of relevance, that is, the relevant word. A result word having the smallest average distance from each word stored in the word storage unit 143 in the word dictionary having the thesaurus structure is selected as a result word.

If there are a plurality of words having the highest degree of association, the result word determination unit 144 determines that the candidate word input first to the result word determination unit 144 (the average candidate rank is the highest). The upper candidate word) is selected as the result word.

In the example of the column of Table 3 (E), the "character" among the four candidate words of the input character string "character" is the distance from the already recognized "pattern" and "recognition" words. Since it is the smallest, it is determined that the degree of association is strong. Therefore, the word “character” is selected as the result word, is output to the output terminal 150, and is simultaneously stored in the word storage unit 143.

The output terminal 150 is a data output terminal for outputting the recognition result to the outside. Via this output terminal 150, the character recognition device of the present invention can be used for other systems (for example, a device for translating using the character recognition result, etc.), a medium for storing the character recognition result, a character recognition result for other places It can be connected to a communication network that communicates with another system or other information processing system.

In the above, the embodiment of the character recognition device of the present invention has been described, but the present invention is not limited to the above-mentioned embodiment, the operation of each component, the processing method by each component, The flow of input / output signals between the constituent components and the number and position of the constituent components can be arbitrarily changed.

For example, in the above embodiment, the character recognition unit 1
10 is configured as shown in FIG. This is because, as the character recognition method, the method of extracting the sub-patterns for each direction as described above, extracting the feature amount indicating the length of the character line from the sub-pattern, and collating the feature amount with the dictionary is used. However, the character recognition method is not limited to this, and other conventionally known methods may be used. Therefore, the configuration of the character recognition unit can be changed to a configuration suitable for the method used.

The word dictionary is not limited to the examples shown in FIGS. Other search methods are possible, items are classified by other methods, additional data is added, words that are related to each other by other methods, etc.
It can be any suitable word dictionary.

In the above embodiment, the unit character string is a word (including a character string in which the same character type is connected). However, it is obvious that the present invention can be applied even when the unit character string is another character string such as a word, a compound word, or a clause.

[0058]

As is apparent from the above description, according to the character recognition device of the present invention, selecting a result character string from one or more candidate unit character strings is one or more to be selected. Can be performed based on the degree of association with another unit character string included in the recognition target document that is the basis for extracting the candidate word character string. Therefore,
Effective post-processing of selecting a unit character string that most closely matches the meaning of a recognition target document from one or more candidate unit character strings can be performed independently of character unit recognition. Moreover, this post-processing is easier than the method of verifying the mutual connection of words by the method of syntax analysis or semantic analysis.

Therefore, since highly accurate recognition can be performed by a simple operation, it is possible to provide a high-performance character recognizing device which can input a large amount of documents at high speed and efficiently.

[Table 1]

[0060]

[0061]

[Table 2]

[0062]

[0063]

[Table 3]

[0064]

[Brief description of drawings]

FIG. 1 is a configuration diagram illustrating an embodiment of a character recognition device of the present invention.

FIG. 2 is a diagram showing a configuration of a character recognition unit included in the character recognition device according to the embodiment.

FIG. 3 is an explanatory diagram of a form example of an example of a document to be recognized.

FIG. 4 is a diagram conceptually showing the structure of a word dictionary unit in the embodiment.

FIG. 5 is a diagram showing a description example of a word dictionary according to the embodiment.

[Explanation of symbols]

S ... Optical signal from document 100: Character recognition device of the embodiment 110: Character recognition unit 120: word area cutout part (unit character string cutout part) 130: Candidate word creation unit (candidate unit character string creation unit) 140: Result word selection unit (result unit character string selection unit) 141: word dictionary (language dictionary) 142: word storage unit (unit character string storage unit) 143: Relevance detection unit 144: Result word determination unit (result unit character string selection unit) 150: Output terminal 200: Report 201: Character entry frame

Claims (5)

[Claims] 1. Document image data is obtained from a document on a medium,
A character recognition unit that recognizes a character string in the document from the document image data to obtain a candidate character string in which each character is one or more candidate characters, and an area corresponding to a unit character string is cut out from the candidate character string. A unit character string cutout unit, a candidate unit character string creation unit that extracts one or more candidate unit character strings from the cutout candidate character strings in each region, and a unit that is a result from the extracted candidate unit character strings. In a character recognition device comprising a result unit character string selection unit for selecting a character string, the result unit character string selection unit includes a language dictionary in which predetermined unit character strings are registered in association with each other, and the result unit character string. A unit character string storage unit that stores a unit character string that is a result previously selected by the selection unit, one or more candidate unit character strings created by the candidate unit character string creation unit, and the unit character string storage unit. Record A degree-of-association detection unit that detects the degree of association with the stored unit character string using the language dictionary, and a degree of relevance detected by the degree-of-association detection unit from the one or more candidate unit character strings. A character recognition device comprising: a result unit character string selection unit that selects a result unit character string based on the result. 2. The character recognition device according to claim 1, wherein the language dictionary has a thesaurus structure language in which predetermined unit character strings are hierarchically classified and the unit character strings are arranged in a tree structure or a network structure. A character recognition device characterized by being a dictionary. 3. The character recognition device according to claim 1, wherein the degree-of-association detection unit includes a unit character string of a superordinate concept and a subordinate concept of a unit character string registered in advance in association with each other in the language dictionary. Character recognition characterized in that the distance to the unit character string is 1, and the distance of the shortest path connecting the candidate unit character string and the unit character string registered in the language dictionary is detected as the degree of association. apparatus. 4. The character recognition device according to claim 1, wherein the unit character string selection unit selects one of the result units stored in the unit character string storage unit among the one or more candidate unit character strings. Select the candidate unit character string with the highest degree of association with the unit character string, or the candidate unit character string with the highest average relevance value as the result unit character string if there are multiple result unit strings. Character recognition device characterized by being a thing. 5. The character recognition device according to claim 1, wherein the unit character string is a connected character string of the same character type,
A character recognition device, which is a character string formed by combining one or more of a word, a compound word, and a clause.